A-lab

AI tools have evolved far beyond simply answering questions like a chatbot. Today, they build websites, edit code, conduct research, generate slides, operate browsers directly, and even execute real actions inside business applications.

But this raises an important question: Are these tools actually worth paying for? And what do real users think of them?

As of May 2026, the most frequently discussed AI tools can be grouped into five categories: general-purpose AI agents like Manus, Anthropic’s Claude and Claude Code, OpenAI’s ChatGPT Agent and Codex, the developer-favorite Cursor, and research-focused tools like Perplexity, Gemini Deep Research, and NotebookLM.

The bottom line: there is no single “best” AI tool that works for everything. Each tool excels at different tasks, and user reactions — both praise and complaints — are quite distinct.

Quick Summary

Claude is still widely praised for coding, writing, and complex reasoning — but complaints about usage limits and inconsistent quality are growing.

OpenAI stands out for versatility, with ChatGPT and Codex covering writing, research, coding, and agent tasks under a single subscription.

Manus sometimes feels like a true AI assistant, but there are significant complaints about credit consumption, customer support, and reliability.

Cursor is widely regarded as the most natural AI coding environment for developers.

Perplexity, Gemini, and NotebookLM are useful for research, but important information should always be verified against original sources.

1. Manus AI: Strong Agent Feel, But Still Risky to Fully Trust

Manus is one of the most talked-about AI agents right now. Rather than simply answering questions, it executes tasks directly — research, website creation, slide generation, design, browser automation, and file analysis. According to official documentation, Manus uses a credit-based system where credits are consumed based on the complexity and resources required for each task.

Manus’s biggest strength is the feeling of “delegating work.” For example, if you ask it to “research this market and compile it into a report,” it goes beyond just answering — it finds sources, builds structure, and delivers a formatted output. Users who appreciate Manus often say it finally feels like AI is doing actual work, not just searching.

However, the downsides are real. The biggest complaint is unpredictable credit consumption. Manus offers free and paid plans, with the Pro plan reportedly starting at $20/month for 4,000 credits. Many users find that a few complex tasks can quickly drain their credits. Another issue is customer support and billing complaints — Trustpilot shows a notable number of negative reviews about subscription cancellations, refunds, and payment issues.

My take: Manus is worth trying for tasks based on publicly available information — blog research, market analysis, simple report drafts, or website brainstorming. However, using it with internal company data, customer information, sensitive financial documents, or production systems still feels risky. Manus is less like a finished AI employee and more like a fast-working but constantly-needs-checking freelance intern.

2. Anthropic Claude / Claude Code: Still a Powerful Choice for Developers and Writers

Claude remains one of the most powerful AI tools in 2026. Claude Code in particular is widely used among developers. Anthropic describes Claude Code as an agentic coding environment that understands codebases, reads files, executes commands, and modifies code.

Claude’s strengths come in three main areas: handling long contexts well, producing natural-sounding writing, and delivering strong results on complex coding and refactoring tasks. Developer communities consistently praise Claude Code for being “great for complex tasks,” “having a strong sense of code structure,” and “feeling more like pair programming than simple autocomplete.”

That said, Claude has significant complaints too. The most prominent issue is the quality degradation controversy. Between March and April 2026, Claude Code users widely reported that performance had suddenly dropped — “it got dumber,” “it’s making unnecessary mistakes.” Anthropic published an official postmortem on April 23, 2026, acknowledging three issues in Claude Code and the Agent SDK, and confirming fixes as of April 20. This incident was an important reminder: even the best AI tools can have their quality fluctuate unexpectedly.

The second major complaint is usage limits. Claude is praised for quality, but heavy users frequently hit restrictions — especially developers running long Claude Code sessions who are sensitive to session limits, token usage, and cost predictability.

My take: Claude remains top-tier for writing, analysis, coding, refactoring, and long document processing. Claude Code can significantly boost productivity when used alongside a developer who reviews its output. But Claude is less a “senior developer replacement” and more a “fast, smart junior developer” — it can point you in the right direction, but humans need to do the final review.

3. OpenAI ChatGPT / Codex / ChatGPT Agent: The Most Versatile Option

OpenAI’s strength is versatility. ChatGPT alone covers writing, translation, research, image generation, data analysis, file handling, coding, and automation. Codex is making a strong resurgence — moving beyond answering code questions to reading codebases, implementing features, fixing bugs, and proposing PR-style results.

User reactions show Codex has a different advantage from Claude Code. Where Claude Code feels like fast, conversational pair programming, Codex is praised for fitting an asynchronous workflow — “assign the task and check the results later.” However, costs can escalate quickly. A Business Insider report shared the case of Every CEO Dan Shipper, who reportedly spent around $13,000 in a single month on personal Codex overages.

ChatGPT Agent is also noteworthy. OpenAI introduces it as an agent mode capable of research, web tasks, slide generation, connector integrations, and browser automation. However, OpenAI openly acknowledges the risks of agent-based AI — particularly prompt injection attacks, where malicious instructions hidden in websites or documents can influence AI behavior.

My take: OpenAI tools are the safest bet for anyone who wants to handle as many tasks as possible with a single subscription. The balance across writing, research, data analysis, coding, and document work is excellent. However, tools that take real actions — like ChatGPT Agent or Codex — require constant supervision. Tasks like sending emails, processing payments, deleting data, handling customer information, or pushing production code must always require human final approval.

4. Cursor: The Most Natural AI Coding Environment for Developers

Cursor is one of the most popular AI coding tools among developers. Its strength lies not in the underlying model, but in the UX — reading, editing, and diffing code, and chatting, all directly within the IDE. Cursor fans commonly say: “It’s convenient to work with AI right inside the code editor,” and “The autocomplete, chat, and file editing flow feels natural.”

For developers who write code every day, Cursor may be the most intuitive tool available — far more natural than copying and pasting code into ChatGPT or Claude. However, Cursor isn’t perfect. For large architectural changes, complex refactoring, or long autonomous tasks, many users run Cursor alongside Claude Code or Codex. And since AI-modified code isn’t always safe, review and testing are non-negotiable.

My take: If you’re a developer, Cursor is worth trying at least once — especially for frontend, web apps, SaaS, personal projects, and rapid prototyping. But Cursor is not an “automatic coding machine.” AI-generated code must be tested, and security or database-related work should always be reviewed by a human.

5. Perplexity, Gemini Deep Research, NotebookLM: Great for Research, But Always Verify

AI research tools are evolving rapidly. Perplexity excels at quickly finding web sources, summarizing them, and showing citations. However, recent user communities have noted that “it’s not as dominant as before” and “Gemini and ChatGPT search have caught up significantly.”

Gemini Deep Research’s strength is its integration with the Google ecosystem. Google explains that Gemini Deep Research can conduct complex research not just from the web, but also from Workspace content like Gmail, Docs, Drive, and Chat. NotebookLM takes a different approach — it excels at organizing and answering questions based on documents, PDFs, links, and materials you provide. It’s less a tool for browsing the entire web and more a “research notebook that helps you understand your own materials.”

My take: For quick web research, Perplexity, ChatGPT, and Gemini are good choices. For organizing your own PDFs and documents, NotebookLM is excellent. For drafting long reports or blog posts, ChatGPT or Claude works best. But always verify important information against original sources — AI research tools can miss context, pull outdated information, or create false confidence in their citations.

The Real Problems People Commonly Report with AI Tools

First: Cost is hard to predict. You think you’re paying $20/month, but the actual cost structure becomes complex due to credits, tokens, overage fees, and plan limits.

Second: Quality is inconsistent. AI tools update fast, but updates don’t always feel like improvements. Like the Claude Code quality controversy, users can suddenly feel “it got worse overnight.”

Third: The more powerful the automation, the bigger the potential accident. In 2026, there was a reported incident where a Claude-based Cursor agent deleted a company’s production database and backups. This illustrates how much risk can emerge when AI agents are given powerful permissions.

Fourth: AI reduces work but creates a new job called “verification.” AI can produce drafts quickly, but checking whether the results are correct, sources are accurate, and code is safe is still entirely a human responsibility. Ultimately, AI tools don’t eliminate work — they change its form.

So Which AI Tool Should You Use?

General work, writing, and document tasks: ChatGPT or Claude. ChatGPT covers a wide range with file analysis, images, data, web search, and agent features. Claude is praised for natural writing tone, long context handling, and analytical responses.

Developers: Using Cursor, Claude Code, and Codex based on purpose is most practical. Cursor is great for daily IDE coding. Claude Code is strong for complex refactoring and structural understanding. Codex is increasingly powerful for async tasks, feature implementation, bug fixes, and code review.

Research and market analysis: Perplexity, ChatGPT, and Gemini for quick web research. Gemini Deep Research if you use Google Workspace heavily. NotebookLM for organizing your own PDFs and materials. The key is not to blindly trust AI summaries.

Automation and AI agents: Manus and ChatGPT Agent offer the experience of “delegating work.” They’re useful for public data research, simple web tasks, slide drafts, and document organization. But be extremely careful when connecting to real accounts, payments, customer data, databases, or internal company systems.

Final Conclusion: In 2026, AI Tools Are About “Controllability,” Not Just “Intelligence”

Every company in the 2026 AI landscape is talking about “AI agents.” AI is no longer just answering questions — it’s moving toward actually executing things. But what users actually feel is slightly different. More important than “How smart is this AI?” are these questions:

• How much can I delegate to this AI?
• Are the costs predictable?
• Can I limit the damage when it makes a mistake?
• Is there a structure in place for me to review the output?

By current standards, AI tools are not perfect employees. They’re more like junior team members — fast and smart, but occasionally doing something unexpected. Used well, they dramatically boost productivity. Used poorly, costs leak, outputs are wrong, and in serious cases, real system accidents can occur.

The best strategy right now is not to hand everything to a single AI tool, but to divide tasks by their nature: writing and general work to ChatGPT or Claude; development to Cursor, Claude Code, or Codex; research to Perplexity, Gemini, or NotebookLM; automation experiments to Manus or ChatGPT Agent.

And the final judgment must always remain with humans. The real AI skill in 2026 is not knowing which tool is the smartest — it’s knowing which tasks to delegate to AI, and which ones still require human review.